Intelligent Market-Making in Artificial Financial Markets

advertisement
Intelligent Market-Making in Artificial Financial
Markets
by
Sanmay Das
A.B. Computer Science
Harvard College, 2001
Submitted to the Department of Electrical Engineering and Computer
Science
in partial fulfillment of the requirements for the degree of
Master of Science in Computer Science and Engineering
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2003
@
Massachusetts Institute of Technology 2003. All rights reserved.
A uthor ...
I..............
Department of Electrical Engineering and Computer Science
April 9, 2003
r-.
Certified by............
......
....................
Tomaso roggio
Eugene McDermott Professor
-Thesis Supervisor
Accepted by.........
.
. .........
Arthur C. Smith
Chairman, Department Committee on Graduate Students
MASSACHUSETTS INSTITUTE
OF TECHNOLOGY
JUL 0 7 2003
I IRPAPIFq
2
Intelligent Market-Making in Artificial Financial Markets
by
Sanmay Das
Submitted to the Department of Electrical Engineering and Computer Science
on April 9, 2003, in partial fulfillment of the
requirements for the degree of
Master of Science in Computer Science and Engineering
Abstract
This thesis describes and evaluates a market-making algorithm for setting prices in
financial markets with asymmetric information, and analyzes the properties of artificial markets in which the algorithm is used. The core of our algorithm is a technique
for maintaining an online probability density estimate of the underlying value of a
stock. Previous theoretical work on market-making has led to price-setting equations for which solutions cannot be achieved in practice, whereas empirical work on
algorithms for market-making has focused on sets of heuristics and rules that lack
theoretical justification. The algorithm presented in this thesis is theoretically justified by results in finance, and at the same time flexible enough to be easily extended
by incorporating modules for dealing with considerations like portfolio risk and competition from other market-makers. We analyze the performance of our algorithm
experimentally in artificial markets with different parameter settings and find that
many reasonable real-world properties emerge. For example, the spread increases in
response to uncertainty about the true value of a stock, average spreads tend to be
higher in more volatile markets, and market-makers with lower average spreads perform better in environments with multiple competitive market-makers. In addition,
the time series data generated by simple markets populated with market-makers using
our algorithm replicate properties of real-world financial time series, such as volatility
clustering and the fat-tailed nature of return distributions, without the need to specify
explicit models for opinion propagation and herd behavior in the trading crowd.
Thesis Supervisor: Tomaso Poggio
Title: Eugene McDermott Professor
3
4
Acknowledgments
First of all, I'd like to thank Tommy for being a terrific advisor, patient when results
don't come and excited when they do. Many others have contributed to this work.
Andrew Lo has been a valuable source of direction. Adlar Kim has been a great
partner in research, always willing to listen to ideas, no matter how half-baked. This
thesis benefited from numerous conversations with Sayan Mukherjee on learning and
Ranen Das, Nicholas Chan and Tarun Ramadorai on finance. Jishnu Das provided
help in getting past a sticking point in the density estimation technique. I'd also
like to thank everyone at CBCL, especially Adlar, Sayan, Tony, Luis, Thomas, Gadi,
Mary Pat, Emily and Casey for making lab a fun and functional place to be, and my
parents and brothers for their love and support.
This thesis describes research done at the Center for Biological & Computational
Learning, which is in the Department of Brain & Cognitive Sciences at MIT and
which is affiliated with the McGovern Institute of Brain Research and with the Artificial Intelligence Laboratory. I was partially supported by an MIT Presidential
Fellowship for 2001-2002. In addition, this research was sponsored by grants from
Merrill-Lynch, National Science Foundation (ITR/IM) Contract No. IIS-0085836 and
National Science Foundation (ITR/SYS) Contract No. IIS-0112991. Additional support was provided by the Center for e-Business (MIT), DaimlerChrysler AG, Eastman
Kodak Company, Honda R&D Co., Ltd., The Eugene McDermott Foundation, and
The Whitaker Foundation.
5
6
Contents
1
Introduction
1.1
13
Background . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . .
13
1.1.1
Market Microstructure and Market-Making . . .
. . . . . . .
14
1.1.2
Artificial Markets . . . . . . . . . . . . . . . . .
. . . . . . .
15
1.1.3
Multi-Agent Simulations and Machine Learning
. . . . . . .
17
1.2
Contributions . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . .
18
1.3
O verview . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . .
19
2 The Market-Making Algorithm
3
21
2.1
Market Microstructure Background . . . . . . . .
21
2.2
Detailed Market Model . . . . . . . . . . . . . . .
23
2.3
The Market-Making Algorithm
. . . . . . . . . .
25
2.3.1
Derivation of Bid and Ask Price Equations
26
2.3.2
Accounting for Noisy Informed Traders . .
27
2.3.3
Approximately Solving the Equations . . .
29
2.3.4
Updating the Density Estimate . . . . . .
30
Inventory Control, Profit Motive and Transaction Prices
33
3.1
A Naive Market-Maker . . . . . . . . . . . . . . . . . . . . .
33
3.2
Experimental Framework . . . . . . . . . . . . . . . . . . . .
34
3.3
Inventory Control . . . . . . . . . . . . . . . . . . . . . . . .
34
3.4
Profit M otive . . . . . . . . . . . . . . . . . . . . . . . . . .
38
3.4.1
38
Increasing the Spread . . . . . . . . . . . . . . . . . .
7
4
3.4.2
Active Learning . . . . . . . . . . . . . . . . . . . . . . . . . .
40
3.4.3
Competitive Market-Making . . . . . . . . . . . . . . . . . . .
42
. . . . . . . . . . . . . . . . . . . . . . . . .
44
3.5
The Effects of Volatility
3.6
Accounting for Jumps
. . . . . . . . . . . . . . . . . . . . . . . . . .
45
3.7
Time Series Properties of Transaction Prices . . . . . . . . . . . . . .
46
51
Conclusions and Future Work
4.1
Sum m ary
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
4.2
Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
8
List of Figures
2-1
An example limit order book . . . . . . . . . . . . . . . . . . . . . . .
22
2-2
Example of the true value over time . . . . . . . . . . . . . . . . . . .
25
2-3
The evolution of the market-maker's probability density estimate with
noisy informed traders (above) and perfectly informed traders (below)
2-4
31
The market-maker's tracking of the true price over the course of the
simulation (left) and immediately before and after a price jump (right)
32
3-1
Step function for inventory control and the underlying sigmoid function 35
3-2
Naive market-maker profits as a function of market volatility without
(above) and with (below) inventory control . . . . . . . . . . . . . . .
3-3
36
Sophisticated market-maker profits as a function of market volatility
without (above) and with (below) inventory control . . . . . . . . . .
36
3-4
Market-maker profits as a function of increasing the spread . . . . . .
39
3-5
Market-maker profits with noisy informed traders and perfectly inform ed traders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
3-6
Standardized log returns over time
. . . . . . . . . . . . . . . . . . .
47
3-7
Leptokurtic distribution of absolute standardized log returns . . . . .
48
3-8
Autocorrelation of absolute returns . . . . . . . . . . . . . . . . . . .
49
3-9
Autocorrelation of absolute returns on a log-log scale for lags of 1-100
49
3-10 Autocorrelation of raw returns . . . . . . . . . . . . . . . . . . . . . .
50
9
10
List of Tables
3.1
Average of absolute value of market-maker's inventory holdings at the
end of a simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2
Correlation between market volatility and market-maker profit for marketmakers with and without inventory control . . . . . . . . . . . . . . .
3.3
. . . . . . . . . . . . . . . . . . . . . . . .
37
Average profit (in cents per time period) for market-makers with and
without randomization in the sampling strategy . . . . . . . . . . . .
3.5
37
Average profit (in cents per time period) for market-makers with and
without inventory control
3.4
36
41
Market-maker profits (in cents per time period) and average number
of trades in simulations lasting 50,000 time steps in monopolistic and
competitive environments
3.6
. . . . . . . . . . . . . . . . . . . . . . . .
Market-maker average spreads (in cents) and profits (in cents per time
period) as a function of the standard deviation of the jump process
3.7
43
.
44
Market-maker average spreads (in cents) and profits (in cents per time
period) as a function of the probability of a jump occurring at any
point in tim e
3.8
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
Average profit (in cents per time period) and loss of expectation for
market-makers using different parameters for recentering the probability distribution
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
46
12
Chapter 1
Introduction
In the last decade there has been a surge of interest within the finance community in
describing equity markets through computational agent models. At the same time,
financial markets are an important application area for the fields of agent-based modeling and machine learning, since agent objectives and interactions tend to be more
clearly defined, both practically and mathematically, in these markets than in other
areas. In this thesis we consider market-making agents who play important roles in
stock markets and who need to optimize their pricing decisions under conditions of
asymmetric information while taking into account other considerations such as portfolio risk. This setting provides a rich and dynamic testbed for ideas from machine
learning and artificial intelligence and simultaneously allows one to draw insights
about the behavior of financial markets.
1.1
Background
The important concepts presented and derived in this thesis are drawn from both
the finance and artificial intelligence literatures. The set of problems we are studying with respect to the dynamics of market behavior has been studied in the market
microstructure and artificial markets communities, while the approach towards modeling financial markets and market-making presented here is based on techniques
from artificial intelligence such as non-parametric probability density estimation and
13
multi-agent based simulation.
1.1.1
Market Microstructure and Market-Making
The detailed study of equity markets necessarily involves examination of the processes
and outcomes of asset exchange in markets with explicit trading rules. Price formation
in markets occurs through the process of trading. The field of market microstructure
is concerned with the specific mechanisms and rules under which trades take place in
a market and how these mechanisms impact price formation and the trading process.
O'Hara [23] and Madhavan [21] present excellent surveys of the market microstructure
literature.
Asset markets can be structured in different ways. The simplest type of market is
a standard double auction market, in which competitive buyers and sellers enter their
prices and matching prices result in the execution of trades [13].
Some exchanges
like the New York Stock Exchange (NYSE) employ market-makers for each stock in
order to ensure immediacy and liquidity. The market-maker for each stock on the
NYSE is obligated to continuously post two-sided quotes (a bid quote and an ask
quote). Each quote consists of a size and a price, and the market-maker must honor a
market buy or sell order of that size (or below) at the quoted price, so that customer
buy and sell orders can be immediately executed. The NYSE employs monopolistic
market-makers. Only one market-maker is permitted per stock, and that marketmaker is strictly regulated by the exchange to ensure market quality. Market quality
can be measured in a number of different ways. One commonly used measure is the
average size of the bid-ask spread (the difference between the bid and ask prices). An
exchange like the NASDAQ (National Association of Securities Dealers Automated
Quotation System) allows multiple market-makers for each stock with less regulation,
in the expectation that good market quality will arise from competition between the
market-makers1 .
Theoretical analysis of microstructure has traditionally been an important part of
the literature. Models in theoretical finance share some important aspects that we use
1A detailed exposition of the different types of market structures is given by Schwartz [261.
14
throughout this thesis. For example, the ability to model order arrival as a stochastic
process, following Garman [14] and Glosten and Milgrom [15], is important for the
derivations of optimal pricing strategies presented here. The basic concepts of how
market-makers minimize risk through inventory control and how this process affects
prices in the market are also used in framing the market-maker's decision problem
and deriving pricing strategies (see, for example, Amihud and Mendelson [1]).
The main problem with theoretical microstructure models is that they are typically
restricted to simple, stylized cases with rigid assumptions about trader behavior.
There are two major alternative approaches to the study of microstructure. These
are the experimental markets approach ([11, 17] inter alia) and the artificial markets
approach ([7, 10, 25] inter alia). The work presented in this thesis falls into the
artificial markets approach, and we briefly review the artificial markets literature.
1.1.2
Artificial Markets
Artificial markets are market simulations populated with artificially intelligent electronic agents that fill the roles of traders. These agents can use heuristics, rules,
and machine learning techniques to make trading decisions. Many artificial market
simulations also use an evolutionary approach, with agents entering and leaving the
market, and agent trading strategies evolving over time. Most research in artificial
markets centers on modeling financial markets from the bottom up as structures that
emerge from the interactions of individual agents.
Computational modeling of markets allows for the opportunity to push beyond the
restrictions of traditional theoretical models of markets through the use of computational power. At the same time, the artificial markets approach allows a fine-grained
level of experimental control that is not available in real markets. Thus, data obtained
from artificial market experiments can be compared to the predictions of theoretical
models and to data from real-world markets, and the level of control allows one to
examine precisely which settings and conditions lead to the deviations from theoretical predictions usually seen in the behavior of real markets. LeBaron [18] provides a
summary of some of the early work on agent-based computational finance.
15
There are two major strands of research on agent-based modeling of financial
markets. The first of these focuses on the emergent properties of price processes
that are generated by the markets. Typically, the goal of research that follows this
approach is to replicate observed properties of financial time series in real markets. For
example, the recent paper of Raberto et al [25] follows this approach, implementing
simple traders who place limit orders, along with a model of opinion propagation
among agents in the Genoa Artificial Stock Market. The results described by Raberto
et al show that their model can capture some features of real financial time series,
such as volatility clustering and the leptokurtic distribution of returns.
Lux [20]
also obtains leptokurtic return distributions in his model, which focuses on chaotic
properties of the dynamical system derived from traders changing between chartist
and fundamentalist trading strategies2 .
The other strand of research in artificial markets focuses more on the algorithms
employed by individual traders. This strand attempts to understand the environments
in which particular strategies are successful, and the resulting implications for market
design.
Examples of research that follow this pattern include the reinforcement-
learning electronic market-maker designed by Chan and Shelton [6], recent work in
the Genoa Artificial Market framework by Cincotti et al [8] that studies long-run
success of trading strategies, and the NASDAQ-inspired simulations of Darley et
al [10].
There is a paucity of work on market-making in the artificial markets literature.
Some simulations of the NASDAQ stock market have been carried out, but none of
them have focused on market-maker behavior or on adaptive agents [10, 4]. With
the exception of the work of Chan and Shelton mentioned above, most research on
market-making has been in the theoretical finance literature, such as the important
paper of Garman [14] which was among the first to explicitly formulate the marketmaker's decision problem. Amihud and Mendelson [1] introduced inventory control
considerations for market-making. Glosten and Milgrom [15] solve the market-maker's
2
Interestingly, Lux does not actually implement a multi-agent simulation, but restricts his model
to a level of simplicity at which he can model the entire market as a system of nonlinear differential
equations.
16
decision problem under information asymmetry. This thesis extends theoretical models of market-making and implements them within the context of our artificial market.
1.1.3
Multi-Agent Simulations and Machine Learning
From the perspective of computer science, both multi-agent based simulation and
machine learning have increased their importance as subfields of artificial intelligence
over the last decade or so. As LeBaron [18] points out, financial markets are one of
the most important applications for agent-based modeling because issues of price and
information aggregation and dissemination tend to be sharper in financial settings,
where objectives of agents are usually clearer. Further, the availability of massive
amounts of real financial data allows for comparison with the results of agent-based
simulations.
In general, work on artificial markets incorporates either learning or evolution as
a means of adding dynamic structure to the markets. In settings where the availability of information is a crucial aspect of market dynamics, adaptive agents who can
incorporate information and learn from market trends become important players. For
example, techniques from classification [22] can be used to predict price movements
for chartist agents, and explicit Bayesian learning can be used by decision-theoretic
agents to incorporate all available information into the decision-making process. Techniques for tracking a moving parameter [5, 3] are useful in estimating the possibly
changing fundamental value of a stock. The price-setting process of market-making
essentially forms a control layer on top of an estimation problem, leading to tradeoffs
similar to the exploration-exploitation tradeoffs often found in reinforcement learning
contexts [28]. Competitive market-making poses its own set of problems that need to
be addressed using game-theoretic analysis and considerations of collaborative and
competitive agent behavior [12].
17
1.2
Contributions
The research described in this thesis serves as a bridge in the literature between the
purely theoretical work on optimal market-making techniques such as the paper of
Glosten and Milgrom [15], which we use for the theoretical underpinnings of this work,
and the more realistic experimental work on market-making that has been carried
out by Chan [7] and Darley et al [10]. We derive an algorithm for price setting that
is theoretically grounded in the optimal price-setting equations derived by Glosten
and Milgrom, and generalize the technique to more realistic market settings. The
algorithm has many desirable properties in the market environments in which we
have tested it, such as the ability to make profits while maintaining a low bid-ask
spread.
The market-making algorithm presented in this thesis is flexible enough to allow
it to be adapted to different settings, such as monopolistic or competitive marketmaking settings, and extended with other modules. We present extensive experimental results for the market-making algorithm and extensions such as inventory control.
We analyze the effects of competition, volatility and jumps in the underlying value
on market-maker profits, the bid-ask spread and the execution of trades.
The data from simulations of markets in which market-makers use the algorithms
developed in this thesis yield interesting insights into the behavior of price processes.
We compare the time series properties of the price data generated by our simulations
to the known characteristics of such data from real markets and find that we are
able to replicate some important features of real financial time series, such as the
leptokurtic distribution of returns, without postulating explicit, complex models of
agent interaction and herd behavior 3 as has previously been done in the literature [20,
25]4.
3Some explicit models of herd behavior are presented in the economics literature by Banerjee [2],
Cont and Bouchaud [9] and Orkan [24] inter alia.
4
1t is worth noting that the true value process can induce behavior (especially following a jump)
similar to that induced by herd behavior through informed traders all buying or selling simultaneously based on superior information. However, the mechanism is a much weaker assumption than
the assumption of explicit imitative behavior or mimetic contagion.
18
1.3
Overview
This thesis is structured as follows. Chapter 2 provides necessary background information on market microstructure, introduces the market model, and derives the
equations for price setting that the main market-making algorithm uses.
It also
presents in detail the cornerstone of the market-making algorithm, a technique for
online probability density estimation that the market-maker uses to track the true
underlying value of the stock.
Chapter 3 describes the practical implementation of the algorithm by taking into
account the market-maker's profit motive and desire to control portfolio risk. This
chapter also presents empirical analysis of the algorithm in various different market
settings, including settings with multiple competitive market-makers, and details the
important time series properties of our model. Chapter 4 summarizes the contributions of this thesis and suggests avenues for future work.
19
20
Chapter 2
The Market-Making Algorithm
2.1
Market Microstructure Background
The artificial market presented in this thesis is largely based on ideas from the theoretical finance literature.
Here we briefly review some of the important concepts.
A stock is assumed to have an underlying true value (or fundamental value) at all
points in time. The price at which the stock trades is not necessarily close to this
value at all times (for example, during a bubble, the stock trades at prices considerably higher than its true value). There are two principal kinds of traders in the
market. Informed traders (sometimes referred to as fundamentalist traders) are those
who know (or think they know) the true value of the stock and base their decisions
on the assumption that the transaction price will revert to the true value. Informed
traders will try to buy when they think a stock is undervalued by the market price,
and will try to sell when they think a stock is overvalued by the market price. Sometimes it is useful to think of informed traders as those possessing inside information.
Uninformed traders (also referred to as noise traders) trade for reasons exogenous to
the market model. Usually they are modeled as buying or selling stock at random
(one psychological model is traders who buy or sell for liquidity reasons). Other models of traders are often mentioned in the literature, such as chartists who attempt to
predict the direction of stock price movement, but we are not concerned with such
'For a detailed introduction to the basic concepts of market microstructure see Schwartz [26].
21
Sell Orders
Buy Orders
Size Price ($) Price ($) Size
x1
23.20
23.28
Y1
x2
23.18
23.30
Y2
X3
23.15
23.00
24.25
Y3
X4
Figure 2-1: An example limit order book
models of trading in this thesis.
There are two main types of orders in stock markets. These are market orders
and limit orders. A market order specifies the size of the order in shares and whether
the order is a buy or sell order. A limit order also specifies a price at which the trader
placing the order is willing to buy or sell. Market orders are guaranteed execution
but not price. That is, in placing a market order a trader is assured that it will
get executed within a short amount of time at the best market price, but is not
guaranteed what that price will be. Limit orders, on the other hand, are guaranteed
price but not execution. That is, they will only get executed at the specified price,
but this may never happen if a matching order is not found.
A double auction market in the context of stocks can be defined as a market in
which limit orders and market orders are present and get executed against each other
at matching prices. The limit orders taken together form an order book, in which the
buy orders are arranged in decreasing order of price, while the sell orders are arranged
in increasing order of price (see figure 2-1 for an example). Orders that match are
immediately executed, so the highest buy order remaining must have a lower price
than the lowest ask order remaining. Market orders, when they arrive, are executed
against the best limit order on the opposite side. So, for example, a market buy order
would get executed against the best limit sell order currently on the book.
Double auction markets are effective when there is sufficient liquidity in the stock.
There must be enough buy and sell orders for incoming market orders to be guaranteed
immediate execution at prices that are not too far away from the prices at which
transactions executed recently. Sometimes these conditions are not met, typically
22
for stocks that do not trade in high volume (for obvious reasons) and immediately
following particularly favorable or unfavorable news (when everyone wants to be either
on the buy or sell side of the market, leading to huge imbalances).
Market-makers are traders designated by markets to maintain immediacy and liquidity in transactions. Market-makers are obligated to continuously post two-sided
quotes (bid (for buying) and ask (for selling) quotes) and honor these quotes. Apart
from providing immediacy and liquidity to order execution, market-makers are also
expected to smooth the transition when the price of a stock jumps dramatically, so
that traders do not believe they received unfair executions, and to maintain a reasonable bid-ask spread. Exchanges with monopolistic market-makers like the NYSE
monitor the performance of market-makers on these categories, while markets like
NASDAQ use multiple market-makers and expect good market quality to arise from
competition between market-makers.
2.2
Detailed Market Model
The market used in this thesis is a discrete time dealer market with only one stock.
The market-maker sets bid and ask prices (Pb and Pa respectively) at which it is
willing to buy or sell one unit of the stock at each time period (when necessary we
denote the bid and ask prices at time period i as Pi and Pa). If there are multiple
market-makers, the market bid and ask prices are the maximum over each dealer's
bid price and the minimum over each dealer's ask price. All transactions occur with
the market-maker taking one side of the trade and a member of the trading crowd
(henceforth a "trader") taking the other side of the trade.
The stock has a true underlying value (or fundamental value) V' at time period
i. All market makers are informed of VO at the beginning of a simulation, but do
not receive any direct information about V after that 2 . At time period i, a single
trader is selected from the trading crowd and allowed to place either a (market) buy
2
That is, the only signals a market-maker receives about the true value of the stock are through
the buy and sell orders placed by the trading crowd.
23
or (market) sell order for one unit of the stock. There are two types of traders in the
market, uninformed traders and informed traders. An uninformed trader will place
a buy or sell order for one unit at random if selected to trade. An informed trader
who is selected to trade knows V' and will place a buy order if V' > P, a sell order
if Vi <Pi and no order if Pz
<
V'
Pj.
In addition to perfectly informed traders, we also allow for the presence of noisy
informed traders. A noisy informed trader receives a signal of the true price W' =
V' + i(O,
Uw)
where i(O, ow) represents a sample from a normal distribution with
mean 0 and variance
4S.
The noisy informed trader believes this is the true value of
the stock, and places a buy order if W' > Pj, a sell order if W' < P' and no order if
PiW' < Pa.
The true underlying value of the stock evolves according to a jump process. At
time i + 1, with probability p, a jump in the true value occurs 3 . When a jump
occurs, the value changes according to the equation V+i = V +cD(0,
-) where cD(0, o-)
represents a sample from a normal distribution with mean 0 and variance a. Thus,
jumps in the value can be more substantial at a given point in time than those in
a unit random walk model such as the one used by Chan and Shelton [6], but the
probability of a change in the true value in our model is usually significantly lower
than the probability of a change in the true value in unit random walk models.
This model of the evolution of the true value corresponds to the notion of the
true value evolving as a result of occasional news items. For example, jumps can be
due to information received about the company itself (like an earnings report), or
information about a particular sector of the market, or even information that affects
the market as a whole. When a jump occurs, the informed traders are placed in
an advantageous position. The periods immediately following jumps are the periods
in which informed traders can trade most profitably, because the information they
have on the true value has not been disseminated to the market yet, and the market
maker is not informed of changes in the true value and must estimate these through
orders placed by the trading crowd. The market-maker will not update prices to the
3
p is typically small, of the order of 1 in 1000 in most of our simulations
24
1004
1003 --
1002 --
1001--
CL 1000
999
998
997
99-
0
1
2
3
4
5
6
7
8
9
Time Prid
10
X 104
Figure 2-2: Example of the true value over time
neighborhood of the new true value for some period of time immediately following a
jump in the true value, and informed traders can exploit the information asymmetry.
2.3
The Market-Making Algorithm
The most important feature of the market-making model presented in this thesis is
that the market-maker attempts to track the true value over time by maintaining a
probability distribution over possible true values and updating the distribution when
it receives signals from the market buy or sell orders that traders place. The true
value and the market-maker's prices together induce a probability distribution on the
orders that arrive in the market. The market-maker's task is to maintain an online
probabilistic estimate of the true value, which is itself a moving target.
Glosten and Milgrom [15] analyze the setting of bid and ask prices so that the
market maker enforces a zero profit condition. The zero profit condition corresponds
to the Nash equilibrium in a setting with competitive market-makers (or, more generally in any competitive price-setting framework [12]). Glosten and Milgrom suggest
that the market maker should set P= E[VISell] and P = E[VBuy]. Our marketmaking algorithm computes these expectations using the probability density function
being estimated.
25
Various layers of complexity can be added on top of the basic algorithm. For
example, minimum and maximum conditions can be imposed on the spread, and an
inventory control mechanism could form another layer after the zero-profit condition
prices are decided. Thus, the central part of our algorithm relates to the density
estimation itself. We will describe the density estimation technique in detail before
addressing other possible factors that market-makers can take into account in deciding
how to set prices. For simplicity of presentation, we neglect noisy informed traders in
the initial derivation, and present the updated equations for taking them into account
later.
2.3.1
Derivation of Bid and Ask Price Equations
In order to estimate the expectation of the underlying value, it is necessary to compute
the conditional probability that V = x given that a particular type of order is received.
Taking market sell orders as the example:
E[VISell] =
j
(2.1)
x Pr(V = xISell) dx
Since we want to explicitly compute these values and are willing to make approximations for this reason, we discretize the X-axis into intervals, with each interval
representing one cent. Then we get:
Vi=Vmax
E[VISell] =
'VPr(V = VISell)
E
%i=Vain
Applying Bayes' rule and simplifying:
/=Vmax
E[VISell
=
Z
A Pr(SellIV = V) Pr(V =
Pr(Sell)
%=Vmin
26
K)
Since P is set by the market maker to E[VSell] and the a priori probabilities of
both a buy and a sell order are equal to 1/2:
V4=Vmax
Pb
=2
E
(2.2)
V Pr(Sell|V = V) Pr(V = V)
14=Vmin
Since Vmin <
P
< Vmax,
Vi=Pb
Pb
Vi=Vmax
=2 E
Vi Pr(SelljV =
)+ 2
) Pr(V =
X4=Vmin
V Pr(SellIV =
) Pr(V =
)
V=Pb+l
(2.3)
The importance of splitting up the sum in this manner is that the term Pr(SelljV =
V ) is constant within each sum, because of the influence of informed traders. An
uninformed trader is equally likely to sell whatever the market maker's bid price. On
the other hand, an informed trader will never sell if V > Pb. Suppose the proportion
of informed traders in the trading crowd is a. Then Pr(SellIV < P) = 1 + la and
Pr(SelljV > P) =
--
a. Then the above equation reduces to:
(VPb
Pb =
2
VVmax
a
(+
a)V Pr(V = V) +
\V=Vmin
1
1
)Vi Pr(V = V)
2-
(2.4)
2
V=Pb+l
Using a precisely parallel argument, we can derive the expression for the marketmaker's ask price:
z%=Pa
Pa = 2
(
i=Vmax
-
a)V Pr(V =Vj)+
(4=Vmin
2.3.2
2 +
a)Vi Pr(V =Vj)
(2.5)
Vi=Pa+1
Accounting for Noisy Informed Traders
An interesting feature of the probabilistic estimate of the true value is that the probability of buying or selling is the same conditional on V being smaller than or greater
than a certain amount. For example, Pr(SelljV
=
, Vi < Pb) is a constant, inde-
pendent of V. If we assume that all informed traders receive noisy signals, with the
27
noise normally distributed with mean 0 and variance a2, and, as before, a is the
proportion of informed traders in the trading crowd, then equation 2.3 still applies.
Now the probabilities Pr(SellIV = V) are no longer determined solely by whether
Vi < P or Vi > P. Instead, the new equations are:
1
(1 - a)- + a Pr(i (0, -w) < (Pb - V))
2
(2.6)
1
Pr(SellIV = V, Vi > Pb) = (1 - a)- + a Pr((O, -w) > (Vi - Pb))
2
(2.7)
Pr(SellIV
=
, Vi
Pb)
=
and:
The second term in the first equation reflects the probability that an informed
trader would sell if the fundamental value were less than the market-maker's bid
price. This will occur as long as W = V + i(0, u-w) < P. Similarly, the second term
in the second equation reflects the same probability, except with the assumption that
V > P.
We can compute the conditional probabilities for buy orders equivalently:
Pr(BuyIV = Vi, Vi < Pa)
=
1
+ a Pr((0, -w) > (P - V))
2
(2.8)
1
)- + a Pr((0, o-w) < (Vi - P))
(2.9)
(1 - a)
and:
Pr(BuyIV = Vi, Vi > Pa) = (1 -
2
We can substitute these conditional probabilities back into both the fixed point
equations and the density update rule used by the market-maker. First of all, com-
28
bining equations 2.3, 2.6 and 2.7, we get:
Pb =
2
(
V=Vmin
--1a + a Pr((0, -w) < (Pb- Vi)))Vi Pr(V =V) +
2 2
V,=Vmax
2
Y
1
(-
- - a + a Pr(ij(0, -w) > (V - Pb)))Vi Pr(V = V)
(2.10)
V,=Pb+l
Similarly, for the ask price:
( - a + a Pr( (0, -w) > (Pa - V))) VPr(V = V) +
Pa = 2
V=Vmin
i=Vmax
2
((2 -
-a+aPr((0,-w)
2
(
-
Pa)))ViPr(V
=
) (2.11)
V=Pa+1
2.3.3
Approximately Solving the Equations
A number of problems arise with the analytical solution of these discrete equations
for setting the bid and ask prices. Most importantly, we have not yet specified the
probability distribution for priors on V, and any reasonably complex solution leads
to a form that makes analytical solution infeasible. Secondly, the values of Vmin and
Vmax are undetermined. And finally, actual solution of these fixed point equations
must be approximated in discrete spaces. We solve each of these problems in turn to
construct an empirical solution to the problem and then present experimental results
in the next chapter.
We assume that the market-making agent is aware of the true value at time 0,
and from then onwards the true value infrequently receives random shocks (or jumps)
drawn from a normal distribution (the variance of which is known to the agent). Our
market-maker constructs a vector of prior probabilities on various possible values of
V as follows.
If the initial true value is V (when rounded to an integral value in cents), then
the agent constructs a vector going from V - 4o- to V + 4c- - 1 to contain the prior
value probabilities. The probability that V = Vo - 4c- + i is given by the ith value
29
in this vector 4 . The vector is initialized by setting the ith value in the vector to
f- 0f "2 1 (0,
-)dx where M is the normal density function in x with specified mean
and variance. The reason for selecting 4o as the range is that it contains 99.9% of
the density of the normal, which we assume to be a reasonable number of entries.
The vector is also maintained in a normalized state at all times so that the entire
probability mass for V lies within it.
The fixed point equations 2.10 and 2.11 are approximately solved by using the
result from Glosten and Milgrom that Pb < E[V] < Pa and then, to find the bid
price, for example, cycling from E[V] downwards until the difference between the left
and right hand sides of the equation stops decreasing. The fixed point real-valued
solution must then be closest to the integral value at which the distance between the
two sides of the equation is minimized.
2.3.4
Updating the Density Estimate
The market-maker receives probabilistic signals about the true value. With perfectly
informed traders, each signal says that with a certain probability, the true value is
lower (higher) than the bid (ask) price. With noisy informed traders, the signal
differentiates between different possible true values depending on the market-maker's
bid and ask quotes. Each time that the market-maker receives a signal about the true
value by receiving a market buy or sell order, it updates the posterior on the value of
V by scaling the distributions based on the type of order. The Bayesian updates are
easily derived. For example, for Vi < Pa and market buy orders:
Pr(V
The prior probability V
=
=
VjBuy) =Pr(BuyjV = V) Pr(V = V )
Pr(Buy)
Vi is known from the density estimate, the prior probability
of a buy order is 1/2, and Pr(BuylV
=
Vi, Vi < Pa) can be computed from equation
2.8. We can compute the posterior similarly for each of the cases.
4
It is important to note that the true value can be a real number, but for all practical purposes
it ends up getting truncated to the floor of that number.
30
0 12
-
-
5 steps
10 steps
15 steps
0.1 -
0.08-
0.062
0.04-
0.02 --
-4
0
1
-2
-1
Standard deviations away from the mean (true value
-3
=
2
-0.56)
4
3
0 .06
.OG
10 steps
- -
0 .05I t
0 .04 -
0 .03 -
0 .02-
0 .01
-4
-3
-2
-1
0
1
Standard deviations away from the mean (true value
2
=
3
4
-0.84)
Figure 2-3: The evolution of the market-maker's probability density estimate with
noisy informed traders (above) and perfectly informed traders (below)
An interesting note is that in the case of perfectly informed traders, the signal only
specifies that the true value is higher or lower than some price, and not how much
higher or lower. In that case, the update equations are as follows. If a market buy
order is received, this is a signal that with probability 1(1 - a) + a
=
1+
>,
V > Pa.
Similarly, if a market sell order is received, the signal indicates that with probability
1+1a ,V < P. In the former case, all probabilities for V = Vi, Vi > P are multiplied
by 1 + ia, while all the other discrete probabilities are multiplied by 1 - (I + !a).
Similarly, when a sell order is received, all probabilities for V = Vi, V < P are
multiplied by 1 + ia, and all the remaining discrete probabilities are multiplied by
1
-
( +
a) before renormalizing.
31
1000
1000.2
-
998
.. . . .,
.i..d..-..
s
.
Price
996-
>
994 -
-
992 -L
999.8
-
999.6 -.
r
.
9900
1800
Ti.1 Period
19
5
.
--
0
005
1
a
15
2
T
25eero
5
3
3.0
4
4
3
3.5
4
4.5
100.
x
-10
0
0.
1810
1820
1830
1840
1810
1820
1830
1840
1810
1820
1830
1840
11800
4
--
1850
1870
188
1898
19D0
1880
1870
1880
1898
1908
1860
1870
Iwo0
1890
19W0
188
Ti m P eriod
185
0.02
0
05
1
1.5
2
2
5
s
1800
1850
Figure 2-4: The market-maker's tracking of the true price over the course of the
simulation (left) and immediately before and after a price jump (right)
These updates lead to less smooth density estimates than the updates for noisy
informed traders, as can be seen from figure 2-3 which shows the density functions
5, 10 and 15 steps after a jump in the underlying value of the stock. The update
equations that consider noisy informed traders serve to smoothly transform the probability distribution around the last transaction price by a mixture of a Gaussian and
a uniform density, whereas the update equations for perfectly informed traders discretely shift all probabilities to one side of the transaction price in one direction and
on the other side of the transaction price in the other direction. The estimates for
perfectly informed traders also tend to be more susceptible to noise, as they do not
restrict most of the mass of the probability density function to as small an area as
the estimates for noisy informed traders.
From figure 2-4 we can see that the market-maker successfully tracks the true value
over the course of an entire simulation. Another interesting feature of the algorithm is
that the bid-ask spread reflects the market-maker's uncertainty about the true value
-
for example, it is typically higher immediately after the true value has jumped.
In the next chapter we present empirical results from applying this algorithm
in different market settings and also extend the basic algorithm for market-making
presented here to take into account other factors like inventory control and profit
motive.
32
Chapter 3
Inventory Control, Profit Motive
and Transaction Prices
3.1
A Naive Market-Maker
At this stage, it is necessary to introduce a simple algorithm for market-making.
There are two main reasons to study such an algorithm. First, it helps to elucidate
the effects of some extensions to the main algorithm presented in the last chapter
(which we shall sometimes refer to as the "sophisticated" algorithm), and second,
it provides a basis for comparison. This naive market-maker "surrounds" the last
transaction price with its bid and ask quotes while maintaining a fixed spread at all
times. At the first time period, the market-maker knows the initial true value and
sets its bid and ask quotes around that price. So, for example, if the last transaction
price was Ph and the market-maker uses a fixed spread 6, it would set its bid and ask
quotes at Ph -
and Ph +
respectively.
Given that we do not consider transaction sizes in this thesis, the above algorithm
is actually surprisingly effective for market-making, as it adjusts its prices upwards or
downwards depending on the kinds of orders entering the market. The major problem
with the algorithm is that it is incapable of adjusting its spread to react to market
conditions or to competition from other market-makers, so, as we shall demonstrate,
it does not perform as well (relatively speaking) in competitive environments or under
33
volatile market conditions as algorithms that take market events into account more
explicitly.
3.2
Experimental Framework
Unless specified otherwise, it can be assumed that all simulations take place in a
market populated by noisy informed traders and uninformed traders.
The noisy
informed traders receive a noisy signal of the true value of the stock with the noise
term being drawn from a Gaussian distribution with mean 0 and standard deviation
5 cents. The standard deviation of the jump process for the stock is 50 cents, and
the probability of a jump occurring at any time step is 0.005. The market-maker is
informed of when a jump occurs, but not of the size or direction of the jump. The
market-maker uses an inventory control function (defined below) and increases the
spread by lowering the bid price and raising the ask price by a fixed amount (this is
done to ensure profitability and is also explained below). We report average results
from 50 simulations, each lasting 50,000 time steps.
3.3
Inventory Control
Stoll analyzes dealer costs in conducting transactions and divides them into three
categories [271. These three categories are portfolio risk, transaction costs and the
cost of asymmetric information. In the model we have presented so far, following
Glosten and Milgrom [15], we have assumed that transactions have zero execution
cost and developed a pricing mechanism that explicitly attempts to set the spread to
account for the cost of asymmetric information.
A realistic model for market-making necessitates taking portfolio risk into account
as well, and controlling inventory in setting bid and ask prices. In the absence of
consideration of trade size and failure conditions, portfolio risk should affect the
placement of the bid and ask prices, but not the size of the spread' [1, 27, 16]. If the
'One would expect spread to increase with the trade size.
34
4~5
3.50.5
--
2
05
10
15
20
2
0
3
Absolue valiue of stock inentory
0
45
50
Figure 3-1: Step function for inventory control and the underlying sigmoid function
market-maker has a long position in the stock, minimizing portfolio risk is achieved by
lowering both bid and ask prices (effectively making it harder for the market-maker
to buy stock and easier for it to sell stock), and if the market-maker has a short
position, inventory is controlled by raising both bid and ask prices.
Inventory control can be incorporated into the architecture of our market-making
algorithm by using it as an adjustment parameter applied after bid and ask prices
have been determined by equations 2.10 and 2.11. An example of the kind of function
we can use to determine the amount of the shift is a sigmoid function. The motivation
for using a sigmoid function is to allow for an initial gradual increase in the impact of
inventory control on prices, followed by a steeper increase as inventory accumulates,
while simultaneously bounding the upper limit by which inventory control can play
a factor in price setting. Of course, the upper bound and slope of the sigmoid can be
adjusted according to the qualities desired in the function.
For our simulations, we use an inventory control function that uses the floor of a
real valued sigmoid function with a ceiling of 5 cents as the integer price adjustment
(in cents).
The step function for the adjustment and the underlying sigmoid are
shown in figure 3-1.
Figure 3-2 is a scatter plot that shows the effects of using the above inventory
control function for a naive market-maker using a 6 value of 8 cents (note that the Y
35
2
i
0
-2
CL -10
500
-12
1000
1500
2000
2500
2000
2500
Dllrno btwen initia ndlast tru aalna
0.8 I
0.4
a
8o
0
aoao
0
a20g
-0.2
a3%O
aa *
C- -0.4
a
e,
a
a
0
-
0
500
1500
1000
Dittatanca betweena initial and last tr alue
Figure 3-2: Naive market-maker profits as a function of market volatility without
(above) and with (below) inventory control
4
r
0
0
-. 0 - 0
-2
0
500
1500
1000
2000
Difference between initial and
2500
3000
2500
3000
last true values
4-
200
0
0
0
0
-2
M
0
500
1500
1000
Differenae
between
2000
initial and last true values
Figure 3-3: Sophisticated market-maker profits as a function of market volatility
without (above) and with (below) inventory control
No IC
158.95
39.46
MM Type
Naive
Sophisticated
IC
7.85
8.99
Table 3.1: Average of absolute value of market-maker's inventory holdings at the end
of a simulation
36
MM Type
Naive
Sophisticated
No IC
-0.9470
-0.7786
IC
-0.1736
-0.3452
Table 3.2: Correlation between market volatility and market-maker profit for marketmakers with and without inventory control
MM Type
Naive
Sophisticated
No IC
-0.1672
0.8372
IC
-0.1033
0.8454
Table 3.3: Average profit (in cents per time period) for market-makers with and
without inventory control
axes are on different scales for the two parts of the figure). Figure 3-3 shows the effects
for a market-maker using the sophisticated algorithm 2 . Table 3.1 shows the average
absolute value of inventory held by the market-maker at the end of each simulation
for the different cases. The figures use the absolute value of the difference between
last true value and initial true value as a proxy for estimating market volatility, as
this difference provides a measure of how much a large inventory could affect profit
for a particular simulation. 500 simulations were run for each experiment, and 70%
of the traders were noisy informed traders, while the rest were uninformed.
The results in figures 3-2 and 3-3 and tables 3.2 and 3.3 demonstrate that without
inventory control, market-maker profits are highly correlated with volatility, and the
inventory control module we have suggested successfully removes the dependence of
profit on volatility without reducing expected profit. The differences in profit for the
inventory control and no inventory control cases are not statistically significant for
either the naive or the sophisticated market-maker. In fact, it is somewhat surprising that average profit is not reduced by inventory control, since adding inventory
control is similar to adding additional constraints to an optimization problem. This
effect could be due to the fact that our algorithm is not in fact performing exact
2
For this experiment, the market-maker was modified to increase the spread beyond the zero
profit condition by lowering the bid price by 3 cents and increasing the ask price by 3 cents. The
motivation for this is to use a profitable market-maker, as will become clear in the next section, and
to perform a fair comparison with a naive market-maker that uses a fixed spread of 8 cents.
37
optimization, and the inventory control module may help to adjust prices in the correct direction in volatile markets. Another interesting fact is that the sophisticated
market-making algorithm is less susceptible to the huge losses that the naive marketmaker incurs in very volatile market environments, even without inventory control.
This suggests that the sophisticated algorithm is adapting to different environments
more successfully than the naive algorithm.
3.4
Profit Motive
The zero-profit condition of Glosten and Milgrom is expected from game theoretic
considerations when multiple competitive dealers are making markets in the same
stock. However, since our method is an approximation scheme, the zero profit method
is unlikely to truly be zero-profit. Further, the market-maker is not always in a perfectly competitive scenario where it needs to restrict the spread as much as possible.
In this section, we investigate some possibilities for increasing the spread to ensure
profitability conditions for the market-making algorithm.
3.4.1
Increasing the Spread
The simplest solution to the problem of making profit is to increase the spread by
pushing the bid and ask prices apart after the zero-profit bid and ask prices have been
computed using the density estimate obtained by the market-making algorithm. The
major effect of this on the density estimation technique is that the signals the marketmaker receives and uses to update its density estimate are determined by transaction
prices, which are in turn determined by the bid and ask prices the market-maker has
set. The precise values of the bid and ask prices are quite important to the sampling
of the distribution on trades induced by the true value.
Figure 3-4 shows the profit obtained by a single monopolistic market-maker in
markets with different percentages of noisy informed traders. The numbers on the X
axis show the amount (in cents) that is subtracted from (added to) the zero-profit bid
(ask) price in order to push the spread apart (we will refer to this number as the shift
38
0-5
in.
709/60%in!
-e- 90% inf
-0.
Amount added to ask price and subtracted from bid price (cents)
Figure 3-4: Market-maker profits as a function of increasing the spread
factor). It is important to note that market-makers can make reasonable profits with
low average spreads - an example is given at the end of the section on competitive
market-making.
With lower spreads, most of the market-maker's profits come from the noise factor
of the informed traders, whereas with a higher spread, most of the market-maker's
profits come from the trades of uninformed traders.
Different percentages of in-
formed traders lead to differently shaped curves. With only 50% of the traders being
informed, the market-maker's profit keeps increasing with the size of the spread. However, increasing the spread beyond a point is counterproductive if there are enough
noisy informed traders in the markets, because then the market-maker's prices are
far enough away from the true value that even the noise factor cannot influence the
informed traders to make trades. With 90% of the traders being informed, a global
maximum (at least for reasonable spreads) is attained with a low spread, while with
70% of the traders being informed, a local maximum is attained with a fairly low
spread, although the larger number of uninformed traders allows for larger profits
with rather large spreads.
Another point worth mentioning is that the market-maker's probability density
estimates tend to be more concentrated with more noisy informed traders in the
markets, because each trade provides more information. This leads to the empirical
39
0.5
R
0-
-- 70/ nois informed
0.t
1
5
2
3
4
6
7
8
Amount added to ask price and subtracted from bid price (cents)
9
10
Figure 3-5: Market-maker profits with noisy informed traders and perfectly informed
traders
results being closer to theoretical predictions. For example, the prices leading to zero
profit for the 70% informed and 90% informed cases fall between the 0 and 1 points
on the X axis, which is close to what one would expect from the theory, whereas with
perfectly informed traders zero profit is not obtained without using a large spread.
Figure 3-5 compares the profits obtained by the market-maker with 70% noisy
informed traders as opposed to 70% perfectly informed traders. In the latter case,
there is no advantage to be gained by having a smaller spread as there is with noisy
informed traders.
However, the market-maker's inability to make any profit even
with a high spread seems surprising. This is partly attributable to the fact that
the point at which the distribution is sampled is more important in the perfectly
informed case because the signals only inform the market-maker of the probabilities
that the true value is greater than or less than the last transaction price, instead
of smoothly morphing points around the last transaction price by a mixture of a
Gaussian distribution and a uniform distribution.
3.4.2
Active Learning
The above observation on how the smoothing effect of noisy informed traders on the
posterior distribution helps to maintain, in some senses, a "good" density estimate
40
Bid/Ask Adjustment
4
4r
Perfectly Informed
-0.7382
-0.5119
Noisy Informed
0.8075
0.5594
Table 3.4: Average profit (in cents per time period) for market-makers with and
without randomization in the sampling strategy
points us in an interesting direction for the case of perfectly informed traders. We
can see from figure 3-5 that, counterintuitively, profits increase slowly with increasing
spread. Since the market-maker sets "controls" at P and Pa, it may not be getting
"good" samples for estimating the probability density for values in the region between
Pb
and Pa, which is the most important part of the distribution, since the true value
probably lies in that range.
One way of dealing with this problem is to occasionally sample from the distribution between P and P by not increasing the spread and just using the zero-profit
condition prices some percentage of the time. We tested the effectiveness of this
method in markets with informed traders constituting 50% of the trading crowd, and
the market-maker decreasing the bid price and increasing the ask price by 4 cents. In
the randomized case, the market-maker increased the spread 60% of the time, leaving
it untouched otherwise. The results from this experiment are shown in table 3.4.
"4r" represents the randomized algorithm, and we can see that it outperforms
the non-randomized version (in fact, it also outperforms versions that use higher
spreads) in markets with perfectly informed traders. The same does not hold true
for markets with noisy informed traders, where the loss incurred by not pushing the
spread up at each opportunity dominates any benefits gained from improving the
density estimate. This is because of the smoother nature of the noisy estimate, as
discussed above. There may be an interesting connection between this behavior and
the "exploration-exploitation" tradeoff as thought of in reinforcement learning [28]
a market-maker willing to sacrifice profits temporarily in order to improve its
estimate can make more profit in the long term. The online nature of the problem
decrees that the probability of sampling more aggressively by not increasing the spread
cannot be smoothly decreased over time, but perhaps more sophisticated algorithms
41
for sampling the important areas of the distribution might help performance even
more in terms of profit in the perfectly informed case. Perhaps some of the ideas
for sampling from the distribution more effectively can be adapted to the noisy case.
This is an interesting direction for future work.
3.4.3
Competitive Market-Making
The most important aspect of competitive market-making within the framework in
which we view it is that market-makers are not guaranteed to execute trades just by
being in the market. Instead the highest bid price quoted by any market-maker and
the lowest ask price quoted by any market-maker become the effective market bid-ask
quotes, and the market-makers compete with each other for trades. A market-maker
who does not make a sufficient number of trades will lose out to a market-maker who
makes substantially more trades even if the latter makes less profit per trade.
This effect is particularly obvious in the first experiment shown in table 3.5. In this
experiment, two market-makers who both use the market-making algorithm presented
in this thesis compete with each other, with the difference that one uses a shift factor
of 2 and the other a shift factor of 3 for increasing the spread after the zero-profit
bid and ask prices have been determined. If neither were using an inventory control
mechanism, the market-maker using a shift factor of 3 would in fact make no trades,
because the market-maker using a shift factor of 2 would always have the inside
quotes for both the bid and the ask. The addition of inventory control allows the
market-maker using a shift factor of 3 to make some trades, but this market-maker
makes considerably less profit than the one using a shift factor of 2. In a monopolistic
environment the market-maker using a shift factor of 3 outperforms the market-maker
using a shift factor of 2 and the difference in magnitude of executed trades is not as
large.
It is interesting to compare two different strategies for market-making in competitive and monopolistic environments. The naive market-making algorithm outperforms
the sophisticated algorithm in a monopolistic framework with 70% of the traders being perfectly informed and the rest uninformed. However, it is outperformed by the
42
70% noisy informed traders
Monopolistic
Competitive
MM Type
Profit # Trades Profit # Trades
Soph (shift = 2) 0.6039
38830 0.6216
39464
Soph (shift = 3) 0.0157
594 0.8655
34873
70% noisy informed traders
Competitive
Monopolistic
MM Type
Profit # Trades
Profit # Trades
Naive (spread = 8) -0.8020
17506 -0.0840
35176
Soph (shift = 3)
0.7687
20341 0.8655
34873
70% perfectly informed traders
Competitive
Monopolistic
MM Type
Profit # Trades
Profit # Trades
Naive (spread = 8) -0.9628
12138 -0.5881
23271
Soph (shift = 3)
-0.6379
16331 -0.8422
27581
Table 3.5: Market-maker profits (in cents per time period) and average number of
trades in simulations lasting 50,000 time steps in monopolistic and competitive environments
sophisticated algorithm in the same environment when they are in direct competition
with each other. Although both are incurring losses, the sophisticated algorithm is
making more trades than the naive algorithm under competition, so the improved performance is not a function of simply not making trades. This is the third experiment
reported in table 3.5. The second experiment in table 3.5 shows the performance
of the two algorithms with 70% of the trading crowd consisting of noisy informed
traders and the remaining 30% consisting of uninformed traders. Again, the presence
of competition severely degrades the naive market-maker's performance without significantly hurting the sophisticated market-maker. This suggests that our algorithm
for market-making is robust with respect to competition.
For a market with 70% of the trading crowd consisting of noisy informed traders
and the remaining 30% consisting of uninformed traders, our algorithm, using inventory control and a shift factor of 1, achieves an average profit of 0.0074 ± 0.0369 cents
per time period with an average spread of 2.2934 ± 0.0013 cents. These parameter
43
100
100
50
50
Shift
1
2
1
2
Spread
2.7366
5.0601
2.2934
4.4466
Profit
-0.7141
-0.1410
0.0074
0.6411
Table 3.6: Market-maker average spreads (in cents) and profits (in cents per time
period) as a function of the standard deviation of the jump process
settings in this environment yield a market-maker that is close to a Nash equilibrium
player, and it is exceedingly unlikely that any algorithm would be able to outperform
this one in direct competition in such an environment given the low spread. It would
be interesting to compare the performance of other sophisticated market-making algorithms to this one in competitive scenarios. Another interesting avenue to explore
is the possibility of adaptively changing the shift factor depending on the level of competition in the market. Clearly, in a monopolistic setting, a market-maker is better
off using a high shift factor, whereas in a competitive setting it is likely to be more
successful using a smaller one. An algorithm for changing the shift factor based on
the history of other market-makers' quotes would be a useful addition.
3.5
The Effects of Volatility
Volatility of the underlying true value process is affected by two parameters. One is
the standard deviation of the jump process, which affects the variability in the amount
of each jump. The other is the probability with which a jump occurs. Table 3.6 shows
the result of changing the standard deviation - of the jump process and table 3.7
shows the result of changing the probability p of a jump occurring at any point in
time. As expected, the spread increases with increased volatility, and profit decreases.
A higher average spread needs to be maintained to get the same profit in more volatile
markets.
44
p
0.005
0.005
0.0001
0.0001
Shift
1
2
1
2
Spread
2.2934
4.4466
2.0086
4.0154
Profit
0.0074
0.6411
0.8269
1.4988
Table 3.7: Market-maker average spreads (in cents) and profits (in cents per time
period) as a function of the probability of a jump occurring at any point in time
3.6
Accounting for Jumps
The great advantage of our algorithm for density estimation and price setting is
that it quickly restricts most of the probability mass to a relatively small region of
values/prices, which allows the market-maker to quote a small spread and still break
even or make profit. The other side of this equation is that once the probability mass
is concentrated in one area, the probability density function on other points in the
space becomes vanishingly small. In some cases, it is not possible to seamlessly update
the estimate through the same process if a price jump occurs. Another problem is
that a sequence of jumps could lead to the value leaving the [-4-, 4o] window used
by the density estimation technique
3
If the market-maker is in some way explicitly informed of when a price jump has
occurred (perhaps the market-maker gets a signal whenever news arrives or may have
arrived, like right before an earnings report is released), although not of the size
or direction of the jump, the problem can be solved by recentering the distribution
around the current expected value and reinitializing in the same way in which the prior
distribution on the value is initially set up. In the "unknown jump" case the problem
is more complicated.
We tested certain simple rules relating to order imbalance
which utilize the fact that the cost of not recentering when a jump has occurred
is significantly higher than the cost of recentering if a jump has not occurred. An
example of such a rule is to recenter when there have been k more buy orders than sell
orders (or vice versa) in the last n time steps. Table 3.8 shows the results obtained
3
For cases with perfectly informed traders the first of these problems is typically not critical, but
since the probabilities away from the expected value still represent a significant probability mass,
estimates can become degraded if the expected value is sufficiently skewed away from the mean.
45
k
n
Known
10
5
10
6
10
7
3
5
Profit
0.8259
0.2327
0.3069
0.2678
-0.4892
Loss of expectation
4.2261
5.5484
5.1544
5.3254
6.8620
Table 3.8: Average profit (in cents per time period) and loss of expectation for marketmakers using different parameters for recentering the probability distribution
using different n and k values, where the loss of the expectation is defined as the
average of the absolute value of the difference between the true value and the marketmaker's expectation of the true value at each point in time. Clearly there is a tradeoff
between recentering too often and not recentering often enough. Although there is
a loss to be incurred by waiting for too long after a price jump to recenter, it can
be even worse to recenter too aggressively (such as the n = 5, k
=
3 case). An
interesting avenue for future work, especially if trade sizes are incorporated into the
model, is to devise a classifier that is good at predicting when a price jump has
occurred. Perhaps there are particular types of trades that commonly occur following
price jumps, especially when limit orders and differing trade sizes are permitted.
Sequences of such trades may form patterns that predict the occurrence of jumps in
the underlying value.
3.7
Time Series Properties of Transaction Prices
Liu et al present a detailed analysis of the time series properties of returns in a
real equity market (they focus on the S&P 500 and component stocks) [19]. Their
major findings are that return distributions are leptokurtic and fat-tailed, volatility
clustering occurs (that is, big price changes are more likely to be followed by big price
changes and small price changes are more likely to be followed by small price changes) 4
and that the autocorrelation of absolute values of returns decays according to a power
4
Liu et al are certainly not the first to discover these properties of financial time series. However,
they summarize much of the work in an appropriate fashion and provide detailed references, and
they present novel results on the power law distribution of volatility correlation.
46
6 -
4
4
-b
6-
0
200
400
600
800
1000
1200
1400
1600
1800
2000
Tim stop
Figure 3-6: Standardized log returns over time
law, and is persistent over large time scales, as opposed to the autocorrelation of raw
returns, which disappears rapidly.
Raberto et al are able to replicate the fat tailed nature of the distribution of returns
and the clustered volatility observed in real markets [25]. However, the Genoa Artificial Stock Market explicitly models opinion propagation and herd behavior among
trading agents in a way that we do not 5 . Nevertheless, our model is also able to replicate the important stylized facts of real financial time series, including the leptokurtic
distribution of returns, clustered volatility and persistence of the autocorrelation of
absolute returns.
A return over a particular time period is defined as the ratio of the prices at which
two transactions occur which are separated by that period in time. In our model, a
one step return is the ratio of the prices at which two successive transactions occur.
We record all transaction prices and assume that the intervals between transactions
are the same. All experiments in this section are in a market with 70% noisy informed
traders, 30% uninformed traders, and a market-maker using a shift factor of 1 (which
results in the market-maker's profit being close to zero). We work with log returns in
this model, where the log return is log P,+1 -log P. Figure 3-6 shows the standardized
log returns over time. Standardized log returns are log returns detrended by the mean
5
A jump in the true value will lead informed traders in our model to make the same decisions on
whether to buy or sell, but not because of imitative behavior among the agents themselves.
47
10*
-Return
-- N(O,1
distribution
10'
10
-
10-2
10'
100
IRel
10
Figure 3-7: Leptokurtic distribution of absolute standardized log returns
and rescaled by the standard deviation. The clustering of volatility and the sharp
tails are evident from figure 3-6. Figure 3-7 demonstrates the leptokurtic nature of
the distribution of returns. The fat tail is evident from the right half of the graph,
where the area being covered by the distribution of returns "pokes out" from the area
covered by a normal distribution. The kurtosis for this experiment was 28.7237.
The other important features of real financial time series that our market also
shows are the long-range persistence of the autocorrelation of absolute returns and
the clustering of volatility (figure 3-8 and figure 3-6 respectively). Interestingly, the
decay of autocorrelation appears to be linear, which is in contrast with the power law
decay observed by Liu et al. If we look more closely, the decay is linear on a log-log
scale for the first 25 lags (indicative of a power law decay) (figure 3-9). The long range
persistence of autocorrelation is an important feature of real financial markets [19],
but in comparison, Raberto et al fail to observe persistence of autocorrelation beyond
80 lags in the Genoa market, and they do not see decay consistent with a power law
at any scale. Real markets, the Genoa market and our artificial market all show quick
decay of the autocorrelation of raw returns (figure 3-10).
48
0.25-
0.2-
015
0.15--
01
0.05
Noise level
00
1
2
3
4
5
Lat
6
oX
7
8
9
10a
Figure 3-8: Autocorrelation of absolute returns
:
10,
. . . . . ..
0
.
. . . . ..
0,
0
0.
0*
10'
10
10,
102
Lag
Figure 3-9: Autocorrelation of absolute returns on a log-log scale for lags of 1-100
49
0.8
0.6-
8
0.40E
0.2
0
5
Lag
10
15
Figure 3-10: Autocorrelation of raw returns
50
Chapter 4
Conclusions and Future Work
4.1
Summary
The major contribution of this thesis is the presentation of an algorithm for marketmaking under conditions of asymmetric information in markets with informed and
uninformed traders. Glosten and Milgrom derive the basic concept of setting bid and
ask prices to be the conditional expectations of the true value given that a sell or buy
order is received, but do not extend the concept beyond toy problems [15]. On the
other side of the spectrum, Chan and Shelton develop a reinforcement learning algorithm for market-making that is fairly complex and attempts to deal with multiple
objectives like profit and inventory control simultaneously, but needs many training episodes and has a hard time approaching profitability, even in markets simpler
than the ones we study here [6]. The price setting equations of our market-making
algorithm are theoretically grounded in the work of Glosten and Milgrom, and the
density estimation technique is essentially explicit Bayesian learning. Modules for
inventory control and for increasing profit by increasing the spread can be added to
the algorithm after solution of the expected value equations for price-setting.
Our market-making algorithm displays many qualities that one would expect from
any reasonable market-maker. It increases the spread when it is more uncertain about
the true value (for example, following a jump in the underlying value) and tends to
maintain a higher spread in more volatile markets. Our market-maker also allows us
51
to gain insights into the structure of simple markets. For example, in markets with
large numbers of noisy informed traders, increasing the spread is counterproductive
beyond a point even in the absence of competition because it no longer allows the
market-maker to make profits from the errant estimates of the noisy informed traders.
In competitive dealer markets, as one would expect, market-makers who execute more
trades tend to benefit even if they make less profit per trade, because their quotes
are on the inside more often.
Simple artificial markets populated by the kinds of trading crowds and marketmakers we describe are capable of replicating some of the important time series phenomena of real financial markets. For example, the leptokurtic distribution of returns
and the persistence of the autocorrelation of absolute returns along with the rapid
decay of the autocorrelation of raw returns are important phenomena in financial time
series [19]. These phenomena are replicated to some extent in the artificial markets
described by Lux [20] and Raberto et al [25] among others, but only with explicit
models of opinion propagation and evolutionary behavior in the trading crowd. The
fact that our model does not need to postulate such behavior, instead relying on
the simple interaction between informed and uninformed traders, may point to an
important underlying regularity of such time series phenomena.
In terms of learning, this thesis describes a nonparametric density estimation algorithm that is very successful in the application domain. The importance of smoothness of the density function for good performance is demonstrated by the far superior
performance of the algorithm with noisy informed traders (it is worth noting that
the gains are not primarily from the noise itself, but are mostly due to the improved
estimation - using the perfectly informed estimates with noisy informed traders does
not lead to performance as good as that achieved using the noisy estimates). The
importance of maintaining a good estimate rather than just increasing the spread is
apparent from the success of the "active learning" algorithm that aggressively samples
from the distribution in the perfectly informed case.
52
4.2
Future Directions
Deriving a model for market-making is just a starting point for much exciting research
in agent-based modeling of financial markets. The research directions we believe to
be particularly interesting in this regard are focused on the learning aspects and on
creating richer market models.
We have briefly touched on how active learning can be used to improve the density
estimate and hence the performance of the market-making algorithm. However, the
method we present is naive and has not been studied in depth. It would be worthwhile
to investigate how to sample the distribution to achieve the best estimate possible
while still maintaining a high enough spread to achieve profit. This is particularly
interesting in the competitive framework where an agent cannot compensate for loss
at one time by raising the spread exorbitantly at another.
Perhaps the most interesting issues will arise in the context of multi-agent learning.
One point of departure is to consider the shift factor and have competitive marketmakers using the same basic algorithm try to learn the optimal shift factor to use
when competing against each other. It is conceivable that this could give rise to
cooperative (or collusive) behavior without the need for explicit communication.
The method of dealing with changes in the underlying value presented in chapter
3 is simplistic.
Perhaps it is possible to learn a classifier either online or offline
that predicts when a jump has occurred (or when the probability of a jump having
occurred is high enough to warrant the cost of recentering the distribution). Lastly in
terms of learning, our model makes many assumptions about the market-maker being
aware of certain parameters like the percentage of informed traders and the standard
deviations of the jump process and the noise terms. What if the market-maker had
to estimate these instead of knowing them up-front'?
There are also some fascinating directions for future work in terms of the market
structure and model. Among these is more detailed examination of the time series
'In preliminary experiments, the profitability and low spread seem fairly robust to the marketmaker using a wrong estimate of the number of informed traders, or of the variance of the noise
factor.
53
properties of returns and an analysis of why they differ from real markets in the
characteristics in which they do differ.
We are also interested in calibrating the
artificial market parameters to real markets. For example, the probability of a jump
or the standard deviation of the jump could be usefully linked to occurrences in
real markets. This would give more meaning to the precise values derived from the
experimental results, like market-maker bid-ask spreads. Finally, it is important to
investigate richer, more complex market-models. The first step in this direction is to
incorporate consideration of different trade sizes. Following this, we would also like
to explore markets with different types of traders, including traders who are capable
of placing limit orders.
54
Bibliography
[1] Y. Amihud and H. Mendelson. Dealership Market: Market-Making with Inventory. Journal of FinancialEconomics, 8:31-53, 1980.
[2] Abhijit V. Banerjee. A Simple Model of Herd Behavior. The Quarterly Journal
of Economics, 107(3):797-817, 1992.
[3] M. Bowling and M.M. Veloso. Multiagent Learning Using a Variable Learning
Rate. Artificial Intelligence, 2002. In press.
[4] M. Brown and V. Darley. The Future of Trading: Biology-based Market Modeling
at Nasdaq. Perspectives in Business Innovation, 4, 1999.
[5] N. Cesa-Bianchi and G. Lugosi. Worst-Case Bounds for the Logarithmic Loss of
Predictors. Machine Learning, 43:247-264, 2001.
[6] N. T. Chan and C. R. Shelton. An Electronic Market Maker. Technical Report
200-005, MIT Artificial Intelligence Laboratory, 2001.
[7] T. Chan. Artificial Markets and Intelligent Agents. PhD thesis, Massachusetts
Institute of Technology, 2001.
[8] S. Cincotti, S.M. Focardi, M. Marchesi, and M. Raberto. Who Wins? Study of
Long-Run Trader Survival in an Artificial Stock Market. Physica A, 2002. In
press.
[9] Rama Cont and Jean-Philipe Bouchaud. Herd Behavior and Aggregate Fluctuations in Financial Markets. Macroeconomic Dynamics, 4:170-196, 2000.
55
[10] V. Darley, A. Outkin, T. Plate, and F. Gao. Sixteenths or Pennies? Observations
from a Simulation of the Nasdaq Stock Market.
In IEEE/IAFE/INFORMS
Conference on Computational Intelligence for FinancialEngineering,2000.
[11] D. Davis and C. Holt. Experimental Economics. Princeton University Press,
Princeton, NJ, 1993.
[12] D.Fudenberg and J.Tirole. Game Theory. MIT Press, Cambridge, MA, 1991.
[13] D. Friedman and J. Rust. The Double Auction Market Institutions, Theories,
and Evidence. Technical report, Santa Fe Institute Studies in the Sciences of
Complexity, 1991.
[14] M. Garman. Market Microstructure. Journal of FinancialEconomics, 3:257-275,
1976.
[15] L.R. Glosten and P.R. Milgrom. Bid, Ask and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders.
Journal of Financial
Economics, 14:71-100, 1985.
[16] S.J. Grossman and M.H. Miller. Liquidity and Market Structure. Journal of
Finance, 43:617-633, 1988.
[17] J. Kagel and A. Roth, editors. Handbook of Experimental Economics. Princeton
University Press, Princeton, NJ, 1995.
[18] B. LeBaron.
Agent-based Computational Finance: Suggested Readings and
Early Research. Journal of Economic Dynamics and Control, 24:679-702, 2000.
[19] Y. Liu, P. Gopikrishnan, P. Cizeau, M. Meyer, C. Peng, and H.E. Stanley. Statistical Properties of the Volatility of Price Fluctuations.
Physical Review E,
60(2):1390-1400, 1999.
[20] T. Lux.
The Socio-Economic Dynamics of Speculative Markets:
Interacting
Agents, Chaos, and the Fat Tails of Return Distributions. Journal of Economic
Behavior and Organization, 33:143-165, 1998.
56
[21] A. Madhavan. Market Microstructure: A Survey. Journal of FinancialMarkets,
pages 205-258, 2000.
[22] Tom M. Mitchell. Machine Learning. WCB/McGraw-Hill, New York, NY, 1997.
[23] M. O'Hara. Market Microstructure Theory. Blackwell, Malden, MA, 1995.
[24] Andre Orlean. Bayesian Interactions and Collective Dynamics of Opinion: Herd
Behavior and Mimetic Contagion. Journal of Economic Behavior and Organization, 28:257-274, 1995.
[25] M. Raberto, S. Cincotti, S.M. Focardi, and M. Marchesi. Agent-based Simulation
of a Financial Market. Physica A, 299:319-327, 2001.
[26] Robert A. Schwartz.
Reshaping the Equity Markets: A Guide for the 1990s.
Harper Business, New York, NY, 1991.
[27] H. Stoll. The Supply of Dealer Services in Securities Markets. Journalof Finance,
33:1133-1151, 1978.
[28] R.S. Sutton and A.G. Barto. Reinforcement Learning: An Introduction. MIT
Press, Cambridge, MA, 1998.
57
Download